Exudation and Diabetic Macular Edema Detection in Retinal Fundus Images with a Public
Thread Rating:
  • 0 Vote(s) - 0 Average
  • 1
  • 2
  • 3
  • 4
  • 5
seminar class
Active In SP

Posts: 5,361
Joined: Feb 2011
15-02-2011, 12:35 PM

Diabetic macular edema (DME) is a common complicationof diabetic retinopathy. In a large scale screeningenvironment the former condition can be assessed by detectingexudates (a type of bright lesions) in fundus images. In thiswork, new methods for the detection exudates and the diagnosisof diabetic macular edema are presented. The methods do notrequire a lesion training set so the need of ground-truthingdata is avoided with great time savings and no risk of humanmistakes. A new publicly available dataset with ground-truthdata containing 169 patients from various ethnic groups andconditions. It is employed to evaluate our algorithms and tocompare the performance of other two exudate segmentationmethods. In all of our tests our exudate detection algorithmperformed best and was employed as the basis for the automaticDME diagnosis. We were able to obtain the exudate segmentationand DME diagnosis with AUC of 0.89 in an average of 4.4 secondsfor each image on an 2.6 GHz platform with an unoptimisedMatlab implementation.I.
DIABETIC retinopathy (DR) is a progressive eye diseasethat currently affects millions of people worldwide. Diabeticmacular edema (DME) is a complication of DR and it isoften the actual cause of vision loss and blindness [1]. DMEis diagnosed because of the swelling of the retina in diabetesmellitus patients due to leaking of fluid from blood vesselswithin the macula. The fluid often leaks from microaneurysms,which are swelled areas of the retinal capillaries. Early diagnosisand treatment are essential to avoid or stop visionloss, but the growing number of diabetic patients threatens tocreate problems in the achievement of recommended screeninglevels. An automatic screening system based on fundus imagescould be the solution to this problem. The research communityhas shown the potentials of such systems and it is activelyworking on their development as it is shown by recent largestudies by Abramoff et al. [2] and Philip et al. [3].It is not possible to directly quantify the retinal thickeningfrom a single fundus image because of the lack of 3Dinformation. However, ophthalmologists infer the presence ofL. Giancardo is with the University of Burgundy and the Oak RidgeNational Laboratory (e-mail: giancardol@ornl.gov).F. Meriaudeau is with the University of Burgundy.T. P. Karnowski and K. W. Tobin Jr. are with the Oak Ridge NationalLaboratory.E. Chaum and Y. Li are with the University of Tennessee Health ScienceCenter. the fluid that causes the retina thickening in diabetic patientsby the presence of lipid deposits. These lipid deposits arecalled exudates. They appear as bright structures with welldefined edges and variable shapes. In this paper we addressthree aspects to the detection of DME: the dataset, the exudatesegmentation and the DME diagnosis.In the medical imaging field, publicly available annotateddatasets of retinal images are a very scarce resource. Weare aware only of five of them: STARE [4], DRIVE [5],DIARETDB1 [6], MESSIDOR [7] and ROC [8]. Each oneof them has different aims, including vessel segmentation,DR diagnosis, microaneurysms localisation. Whenever a singlecommon dataset was employed by different research groupsthe advantages and disadvantages of each method proposedwere easily measurable and comparable as shown in [8].The only dataset containing manually segmented exudatesis DIARETDB1, but unfortunately the majority of methodsfor exudate detection found in the literature were tested onindependent datasets with various characteristics, and withvery different evaluation methods (see Table IV). This isvery problematic, especially considering that the appearance ofhuman retina significantly changes from person to person. Tocontribute to the development of ever better image analysismethods, we have made our dataset public. This datasetcontains high quality fundus images of patients from differentethnic backgrounds. In addition to the images, a manuallyproduced ground-truth lesion map and other meta-data areprovided. The reason why the DIARETDB1 is not employedis because of the lack of ethnic variability in it. In fact, 88of its 89 images show the pigmentation typical of Caucasians,which represents only a small percentage of the populationrequiring mass screening in rural areas of USA and the world.The approaches to exudates segmentation presented in theliterature can be roughly divided into four different categories:thresholding methods base the exudate identification of aglobal or adaptive grey level analysis. A first attempt waspresented in [9] and only recently a more sophisticated methodbased on image normalisation and distribution analysis waspresented in [10]. Region growing methods segment the imagesusing the spacial contiguity of grey levels; a standard regiongrowing approach is used in [11], which is very computationallyexpensive by being employed to the whole image. In [12]the computational issues are limited by employing edge detec tion to limit the size of regions. Morphology methods employgreyscale morphological operators to identify all structureswith predictable shapes (such as vessels), these are removedfrom the image so that exudates can be identified [13], [14].Classification methods build a feature vector for each pixel orpixel cluster, which are then classified by employing a machinelearning approach into exudates or not exudates [15]–[17] oradditional types of bright lesions [18] (such as drusens andcotton wool spots).In this paper we present two variations of a new exudatesegmentation method that falls into the category ofThresholding methods. Hence, the methods do not require anymanual segmented lesion map for training to achieve goodperformance. In order to reduce the problem of noise we havedeveloped a new way to normalise the fundus images. Wedirectly compare our method with an implementation of amorphology based technique [14] and another thresholdingbased technique [10]. Since all the algorithms published in theliterature present excellent results, we selected the techniquesthat were tested with the largest datasets in the respectiveclass. For our current work we chose not to consider regiongrowing and classification methods because the former iscomputationally expensive and the latter requires training ona lesion by lesion basis, which lead to some issues as isillustrated in section III.Finally, we present a new method for the diagnosis of DME.The method presented here does not attempt to grade DMEusing a precise scale like the one presented by Gangnon etal. [19] or the three levels of International Clinical DMESeverity Scale, but rather provides an automated “DME/noDME” grading. This permits the early detection of DME asopposed to any detection of clinically significant DME. Weapproach the detection of DME with a hybrid technique. Wedo use the automatic exudate segmentation to train a classifierable to recognise patients with DME, but we do not try todistinguish true lesions from false lesions; rather, we extracta feature vector on the image as a whole by employing theinitial segmentation as a map. This approach allows us to havethe benefits of the lesion segmentation without any need forthe error prone manual lesion segmentation by retina experts(although our dataset provides one). Furthermore, the lesionsegmentation will allow clinicians to visually understand thereasoning behind the automated diagnosis, which is a crucialstep for the acceptance of such automatic system in a clinicalenvironment.This paper begins with a description of the compiled dataset,its statistical distribution and how it was manually segmentedin Section II; Section III reviews two segmentation techniquesfrom the literature which do not require any direct training onmanually segmented lesions; Section IV introduces two newexudate segmentations techniques that again do not requireany explicit training phase; Section V shows the results obtainedwith the four automatic lesion segmentation techniquesimplemented; Section VI describes an automatic method todiagnose DME using the segmentation previously obtained;Section VII presents the final results for the diagnosis; finally,section VIII discusses and concludes the paper.

download full report


Important Note..!

If you are not satisfied with above reply ,..Please


So that we will collect data for you and will made reply to the request....OR try below "QUICK REPLY" box to add a reply to this page

Quick Reply
Type your reply to this message here.

Image Verification
Please enter the text contained within the image into the text box below it. This process is used to prevent automated spam bots.
Image Verification
(case insensitive)

Possibly Related Threads...
Thread Author Replies Views Last Post
  disease detection using biorobotics pdf jaseelati 1 648 19-09-2015, 05:42 PM
Last Post: Guest
  Report on Automatic Segmentation of Digital Images Applied in Cardiac Medical Images study tips 0 662 01-03-2013, 01:05 PM
Last Post: study tips
  R Peak Detection in Electrocardiogram Signal Based on an Optimal Combination study tips 0 639 25-02-2013, 12:15 PM
Last Post: study tips
  Detection of Small Bowel Polyps and Ulcers in Wireless Capsule Endoscopy Videos pdf project girl 0 541 17-11-2012, 02:42 PM
Last Post: project girl
  Denoising, Segmentation and Characterization of Brain Tumor from Digital MR Images project girl 0 671 06-11-2012, 05:57 PM
Last Post: project girl
  An Artificial Neural Network for Detection of Biological Early Brain Cancer seminar ideas 1 1,350 05-10-2012, 11:51 AM
Last Post: seminar tips
  Event Detection seminar flower 0 1,673 29-09-2012, 10:31 AM
Last Post: seminar flower
  SPINTRONIC SCANNER FOR CANCER DETECTION seminar ideas 1 1,325 13-08-2012, 10:10 AM
Last Post: seminar ideas
  A Modified Method for Speckle Noise Removal in Ultrasound Medical Images project uploader 0 546 23-07-2012, 03:47 PM
Last Post: project uploader
Last Post: seminar ideas